-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathtrain.py
182 lines (154 loc) · 8.73 KB
/
train.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
import os
import cv2
import numpy as np
import tensorflow as tf
import cpm_model
import tf_utils, cpm_utils
'''
Parameters
'''
FLAGS = tf.app.flags.FLAGS
tf.app.flags.DEFINE_string('tfr_data_dir',
default_value=['./inputs/cpm_aichallenger_dataset.tfrecords'],
docstring='Training data tfrecords')
tf.app.flags.DEFINE_string('pretrained_model',
default_value='./models',
docstring='pretrained mode')
tf.app.flags.DEFINE_integer('input_size',
default_value=128,
docstring='input image size')
tf.app.flags.DEFINE_integer('heatmap_size',
default_value=32,
docstring='output heatmap size')
tf.app.flags.DEFINE_integer('stages',
default_value=6,
docstring='CPM stages number')
tf.app.flags.DEFINE_integer('joints_num',
default_value=14,
docstring='joints number')
tf.app.flags.DEFINE_integer('batch_size',
default_value=32,
docstring='training batch size')
tf.app.flags.DEFINE_integer('training_iterations',
default_value=300000,
docstring='max training iterations')
tf.app.flags.DEFINE_float('learning_rate',
default_value=0.0001,
docstring='initial learning rate')
tf.app.flags.DEFINE_float('lr_decay_rate',
default_value=0.9,
docstring='learning rate decay rate')
tf.app.flags.DEFINE_integer('lr_decay_step',
default_value=2000,
docstring='learning rate decay steps')
tf.app.flags.DEFINE_string('saved_model_path',
default_value='cpm_i%s*%s_o%s*%s_s%s',
docstring='saved model path')
tf.app.flags.DEFINE_string('saved_log_path',
default_value='cpm_i%s*%s_o%s*%s_s%s',
docstring='saved log path')
tf.app.flags.DEFINE_string('saved_name',
default_value='cpm_i`%s*%s_o%s*%s_s%s',
docstring='saved results name mode')
def main(argv):
result_name = FLAGS.saved_log_path % (str(FLAGS.input_size), str(FLAGS.input_size),
str(FLAGS.heatmap_size), str(FLAGS.heatmap_size), str(FLAGS.stages))
trained_model_dir = os.path.join('.', 'results', 'models', result_name)
log_dir = os.path.join('.', 'results', 'logs', result_name)
min_loss = 300
batch_x, batch_y, batch_x_ori = tf_utils.read_batch_cpm(FLAGS.tfr_data_dir, FLAGS.input_size,
FLAGS.joints_num, batch_size=FLAGS.batch_size)
input_x = tf.placeholder(dtype=tf.float32, shape=(FLAGS.batch_size, FLAGS.input_size, FLAGS.input_size, 3),
name='input_x')
input_y = tf.placeholder(dtype=tf.float32,
shape=(FLAGS.batch_size, FLAGS.heatmap_size, FLAGS.heatmap_size, FLAGS.joints_num),
name='input_heatmap')
model = cpm_model.CPM_Model(FLAGS.stages, FLAGS.joints_num+1)
model.build_model(input_x, FLAGS.batch_size)
model.build_loss(input_y, FLAGS.learning_rate, FLAGS.lr_decay_rate, FLAGS.lr_decay_step)
print('====Model Build====\n')
'''Training'''
with tf.Session() as sess:
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord)
tf_writer = tf.summary.FileWriter(log_dir, sess.graph, filename_suffix=result_name)
saver = tf.train.Saver(max_to_keep=None)
init_op = tf.global_variables_initializer()
sess.run(init_op)
if FLAGS.pretrained_model is not None:
saver.restore(sess, FLAGS.pretrained_model)
for variable in tf.trainable_variables():
with tf.variable_scope('print weights', reuse=True):
var = tf.get_variable(variable.name.split(':0')[0])
print(variable.name, np.mean(sess.run(var)))
while True:
batch_x_tf, batch_y_tf = sess.run([batch_x, batch_y])
# warp training images
for img_num in range(batch_x_tf.shape[0]):
deg1 = (2 * np.random.rand() - 1) * 50
deg2 = (2 * np.random.rand() - 1) * 50
batch_x_tf[img_num, ...] = cpm_utils.warpImage(batch_x_tf[img_num, ...], 0, deg1, deg2, 1, 30)
batch_y_tf[img_num, ...] = cpm_utils.warpImage(batch_y_tf[img_num, ...], 0, deg1, deg2, 1, 30)
batch_y_tf[img_num, :, :, FLAGS.joints_num] = np.ones(shape=(FLAGS.input_size, FLAGS.input_size))\
- np.max(batch_y_tf[img_num, :, :, FLAGS.joints_num],
axis=2)
# Recreate heatmaps
heatmaps_tf = cpm_utils.make_gaussian_batch(batch_y_tf, FLAGS.heatmap_size, 3)
stage_losses, total_loss, _, summary, current_lr, stage_heatmaps, global_step =\
sess.run([model.stage_loss,
model.total_loss,
model.train_op,
model.merged_summary,
model.lr,
model.stage_heatmaps,
model.global_step],
feed_dict={input_x: batch_x_tf,
input_y: heatmaps_tf})
tf_writer.add_summary(summary, global_step)
if global_step%50 == 0:
demo_img = batch_x_tf[0] + 0.5
demo_stage_heatmaps = []
for stage in range(FLAGS.stages):
demo_stage_heatmap = stage_heatmaps[stage][0, :, :, 0:FLAGS.num_of_joints].reshape(
(FLAGS.heatmap_size, FLAGS.heatmap_size, FLAGS.num_of_joints))
demo_stage_heatmap = cv2.resize(demo_stage_heatmap, (FLAGS.input_size, FLAGS.input_size))
demo_stage_heatmap = np.amax(demo_stage_heatmap, axis=2)
demo_stage_heatmap = np.reshape(demo_stage_heatmap, (FLAGS.input_size, FLAGS.input_size, 1))
demo_stage_heatmap = np.repeat(demo_stage_heatmap, 3, axis=2)
demo_stage_heatmaps.append(demo_stage_heatmap)
demo_img_heatmap = heatmaps_tf[0, :, :, 0:FLAGS.num_of_joints].reshape(
(FLAGS.heatmap_size, FLAGS.heatmap_size, FLAGS.num_of_joints))
demo_img_heatmap = cv2.resize(demo_img_heatmap, (FLAGS.input_size, FLAGS.input_size))
demo_img_heatmap = np.amax(demo_img_heatmap, axis=2)
demo_img_heatmap = np.reshape(demo_img_heatmap, (FLAGS.input_size, FLAGS.input_size, 1))
demo_img_heatmap = np.repeat(demo_img_heatmap, 3, axis=2)
if FLAGS.stages > 4:
upper_img = np.concatenate((demo_stage_heatmaps[0], demo_stage_heatmaps[1], demo_stage_heatmaps[2]),
axis=1)
blend_img = 0.5 * demo_img_heatmap + 0.5 * demo_img
lower_img = np.concatenate((demo_stage_heatmaps[FLAGS.stages - 1], demo_img_heatmap, blend_img),
axis=1)
demo_img = np.concatenate((upper_img, lower_img), axis=0)
cv2.imshow('current heatmap', (demo_img * 255).astype(np.uint8))
cv2.waitKey(1000)
else:
upper_img = np.concatenate((demo_stage_heatmaps[FLAGS.stages - 1], demo_img_heatmap, demo_img),
axis=1)
cv2.imshow('current heatmap', (upper_img * 255).astype(np.uint8))
cv2.waitKey(1000)
print('##========Iter {:>6d}========##'.format(global_step))
print('Current learning rate: {:.8f}'.format(current_lr))
for stage_num in range(FLAGS.stages):
print('Stage {} loss: {:>.3f}'.format(stage_num + 1, stage_losses[stage_num]))
print('Total loss: {:>.3f}\n\n'.format(total_loss))
# Save models
if total_loss <= min_loss:
min_loss = total_loss
saver.save(sess=sess, save_path=trained_model_dir, global_step=global_step)
print('\nModel checkpoint saved...\n')
# Finish training
if global_step == FLAGS.training_iterations:
break
coord.request_stop()
coord.join(threads)
print('Training done.')